On February 24, 2026, X restricted the POST /2/tweets endpoint so API apps can no longer post replies unless the original author explicitly mentions or quote-posts the replier (Tekedia, 2026). Overnight, a generation of reply bots stopped working — and the question "is AI reply automation safe on X?" got a sharper answer than it had in 2024.
Short version: fully automated replies are no longer safe — the API won't let you, and the ones that bypass the API via browser automation get caught. AI-assisted replies, where the human approves every send, are still compliant and still scale. The line between those two is where every policy fight in 2026 happens.
Key Takeaways
- X's Feb 2026 API change blocks programmatic replies on Free, Basic, Pro, and Pay-Per-Use tiers unless the author summons the replier (X Developers, 2026).
- Fully automated replies — via API or scraped browser sessions — carry real shadowban and suspension risk in 2026.
- AI-assisted replies with human review remain compliant and are the only pattern that still scales.
- Detection thresholds to stay under: ~30 replies/hour, ~100 likes/hour, ~50 retweets/hour, 2,400 posts/day (OpenTweet, 2026).
What Does X's Policy Actually Say About Automated Replies?
X's automation rules explicitly prohibit "automating reply functions to reach many users on an unsolicited basis" and call out sending automated replies based on keyword searches alone (X Help Center, 2025). The line is intent: targeted replies the recipient asked for are fine; keyword-triggered spray-and-pray replies are not.
Three categories matter when you read the policy:
- Allowed: scheduling your own posts through an OAuth-authorized tool, posting AI-generated content you wrote, running a clearly labeled bot account, replying manually even if an AI drafted the text.
- Conditional: automated replies to users who have "requested or clearly indicated an intent to be contacted" (X Help Center, 2025) — think a support bot that only triggers on a specific mention or hashtag opt-in.
- Prohibited: automated follow/unfollow, auto-like, auto-retweet, auto-DM, bulk unsolicited replies, browser automation that uses stolen or shared session cookies, any tool posting at volumes that trip spam detection.
Don't AI-generated posts themselves get banned? No — X's policy doesn't require disclosure of AI-generated content, and the platform has been explicit that using ChatGPT, Claude, or Gemini to write tweets is not a violation by itself. What trips the rule is unsolicited distribution — the reply volume, not the drafting method.
X removed over 1.16 million abuse-related posts in the second half of 2024 and reported a 19% decline in spam across its network (X Global Transparency Report, 2025). Suspensions dropped by roughly 164,000 versus the prior period — cleaner signal, tighter enforcement. The platform is getting better at this, not worse.
What Changed in February 2026? (The API v2 Restriction)
On 24 February 2026, X rolled out a live API v2 change restricting programmatic replies on Free, Basic, Pro, and Pay-Per-Use tiers (Tekedia, 2026). Replies via POST /2/tweets are now rejected unless the original post's author has @mentioned the replying account or quoted one of its posts. Enterprise and Public Utility apps are exempt. Manual user-typed replies are unaffected.
Behavior Status after 24 Feb 2026 Programmatic reply with no prior @mention/quote Blocked (Free/Basic/Pro/PPU) Programmatic reply after author @mentions replier Allowed Programmatic reply after author quotes replier Allowed Enterprise / Public Utility app replies Allowed (exempt) Scheduled original posts via API Allowed Manual reply typed by a logged-in user Allowed Browser automation using shared session cookies Prohibited (pre-existing) Source: X Developers announcement, 2026.
Here's the part most guides miss: the February change didn't kill AI on X — it killed server-initiated AI replies. A tool that runs in your own browser session, that you click to send, that holds an OAuth token scoped to your personal account, can still draft and submit replies. The reason it still works isn't a loophole — it's that from X's side, a logged-in person clicking "Post" looks exactly like a logged-in person clicking "Post." Because it is.
The tools that died overnight were server-side bot farms generating keyword-triggered replies across thousands of accounts through the API. The tools that kept running were user-extension patterns — one user, one session, one click per reply — even when the text came from an AI draft.
How Does X Detect Reply Bots?
X's detection stack layers three signals: behavioral (per-account velocity), graph (network patterns), and content (near-duplicate text across accounts) (OpenTweet, 2026). Crossing any one of them aggressively can trigger a lock in under 48 hours. The behavioral layer is where most individual accounts get flagged first.
Reported hourly thresholds circulating in 2026 — above which detection sharply increases:
- More than 30 replies per hour across your account (Tweet Archivist, 2026).
- More than 100 likes per hour.
- More than 50 retweets per hour.
- More than 400 follows per day or any follow/unfollow pattern that repeats.
The platform-wide ceiling is 2,400 posts per day, with a rolling ~50-post cap over any 30-minute window (tendX, 2026). Replies count against the same bucket as original posts and quote-posts.
The graph layer is the one most people underestimate. "Reply networks" — clusters of accounts that always reply to each other to pump engagement — get detected through interaction graphs, posting-time correlation, and account-creation metadata (Tweet Archivist, 2026). One account getting flagged in a detected network can suspend the whole cluster. If you're running AI replies across five client accounts that all interact with each other, that's a graph pattern.
Content similarity is the third layer. Even if every reply is AI-written and unique-looking to a human, near-duplicate phrasing across accounts — the classic "Great insight! Totally agree 🚀" pattern rewritten 50 ways — still matches. Temperature and output diversity in the underlying LLM matter more than most operators think.
What Happens If You Get Caught? (Shadowban, Lockouts, Suspension)
Penalties on X escalate in stages, and most first-time offenses resolve in under two weeks — if the underlying behavior stops. According to data compiled by Hootsuite's 2025 Social Visibility Report, 68% of creators have experienced sudden engagement drops linked to algorithmic restrictions (OpenTweet, 2026). Most of those drops are shadowbans, not full suspensions.
Restriction type Typical duration What it does Search suggestion ban 24-72 hours Username stops autocompleting in search Reply deboost 48-96 hours Replies hidden behind "show more" in threads Ghost ban (full filter) 5-14 days Tweets invisible to non-followers Repeat-offender escalation 14+ days or permanent Can roll into account suspension Source: OpenTweet shadowban guide, 2026; Pixelscan, 2025.
The timeline matters more than the category. The clock on a shadowban doesn't start when the throttle kicks in — it starts when you actually stop the trigger behavior. Keep auto-replying through a deboost and the 48-hour window resets every day.
We see the same pattern repeatedly in the reply-automation category: an account that's been running fully-automated keyword replies hits a deboost, the operator doesn't realize it for a week because their reply volume looks fine in the tool's dashboard, and by the time they investigate, the ghost-ban window has extended. The tools that survive 2026 aren't the ones optimizing for volume — they're the ones optimizing for sustainable per-account posture.
Full account suspension is rarer but not rare. Anecdotally, suspensions cluster around repeat offenses, multi-account coordination, or use of tools that authenticate via shared session cookies rather than OAuth. The single fastest path to a permanent ban is still paying for a third-party service that logs into your account on a headless browser — that's explicitly prohibited under X's automation rules (X Help Center, 2025).
Can AI-Assisted Replies Be Safe? (The Compliant Pattern)
Yes — AI-assisted replies are compliant when three conditions hold: the user authenticates through X's official OAuth flow, the user reviews every reply before it's sent, and the per-account volume stays under detection thresholds (X Help Center, 2025). That pattern hasn't been broken by any 2025-2026 policy change and isn't projected to be.
The five-signal compliance checklist for any AI reply tool you evaluate in 2026:
- OAuth authentication, not session cookies. If the tool asks for your X password instead of redirecting you to X's permission screen, walk away.
- Human-in-the-loop on every send. AI drafts, human clicks. "Auto-post after 30 seconds unless you cancel" is not human-in-the-loop — it's deferred automation.
- Per-account rate ceilings. The tool caps you below detection thresholds by default and makes it visible when you're getting close.
- Output diversity. Reply drafts should vary in structure, length, and vocabulary — not five templates rewritten 1,000 ways.
- No cross-account coordination. If you run multiple accounts for clients, the tool should treat each session as fully independent and not share content graphs across them.
There's a conversion reason to keep the human in the loop too, not just a policy one. Replies that convert to profile clicks and follows tend to include a specific detail from the parent post — a name, a number, a contrarian take. AI drafts are a starting point, but the 10-15 seconds a human spends editing the draft is where the personalization lives. Tools that remove the human step optimize for volume at the exact place where quality compounds.
Is the slow-and-review pattern competitive with pure automation? At the per-reply level, no. At the per-account level, over a three-month window, yes — because the compliant accounts are still alive in month three. The ones posting 200 automated replies a day in month one are deboosted by week two.
Frequently Asked Questions
Are bot accounts allowed on X in 2026?
Yes, if they're clearly labeled as automated in the bio and comply with X's automation rules (X Help Center, 2025). Weather bots, news-headline accounts, and utility bots are fine. Undisclosed bots pretending to be humans are a policy violation and a common suspension trigger.
Do I need to disclose that I used AI to write my tweets?
No. X doesn't require disclosure of AI-generated content, and using ChatGPT, Claude, or Gemini to draft tweets isn't a violation on its own (OpenTweet, 2026). What matters is the distribution pattern — unsolicited bulk replies get flagged regardless of whether a human or AI wrote them.
What rate limits should I stay under to avoid detection?
Stay below ~30 replies per hour per account, ~100 likes per hour, and ~50 retweets per hour, with a hard ceiling of 2,400 posts per day (tendX, 2026). Replies, quotes, and reposts share the same daily bucket. The rolling 30-minute cap is ~50 posts, which matters more for burst behavior.
Are Taplio, Tweet Hunter, and Typefully safe to use in 2026?
Tools that authenticate via X's official OAuth flow and scope themselves to scheduling or drafting — Typefully, Buffer, Hootsuite, and similar — remain compliant (OpenTweet, 2026). Any tool's reply-automation features need checking against the Feb 2026 API change. Ask each vendor whether their reply flow is OAuth + user-click or server-initiated — the answer matters.
How long does an X shadowban last?
Typical durations: 24-72 hours for search-suggestion bans, 48-96 hours for reply deboosts, 5-14 days for ghost bans (OpenTweet, 2026). The window resets from the day you stop the triggering behavior, not from the day the throttle started. Repeat offenders can see durations roll into weeks or tip into permanent suspension.
So — Is AI Reply Automation Safe on X?
Fully automated reply bots are not safe in 2026 — the API won't run them for most developers, and the bypass paths get caught. AI-assisted replies with user authentication and human review are safe and are the only pattern that compounds past month one. The compliant shape of the workflow is: AI drafts, you edit, you click, X sees a logged-in human posting a personalized reply at human pace.
If you want to run that pattern at scale without crossing any of the 2026 thresholds, that's what Commeta's Reply Guy is built for — OAuth authentication, per-account rate caps, AI drafts with one-click review, no cross-account coordination. Try it free at commeta.app and see what compliant AI replies look like at 100 a day.